Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition
Authors: Hao Helen Zhang, Zhan Zhuang, Xuehao Wang, Xiaodong Yang, Yu Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on six HAR benchmarks demonstrate that Mo PFormer not only outperforms state-of-the-art methods but also successfully generalizes across multiple datasets. |
| Researcher Affiliation | Academia | Hao Zhang1 , Zhan Zhuang1,2 , Xuehao Wang3, Xiaodong Yang4, Yu Zhang1 1Southern University of Science and Technology, 2City University of Hong Kong 3Zhejiang University, 4Institute of Computing Technology, Chinese Academy of Sciences |
| Pseudocode | No | The paper describes the methodology in prose and illustrates the architecture through figures, but no explicit pseudocode or algorithm blocks are present. |
| Open Source Code | No | Code will be provided soon. |
| Open Datasets | Yes | We conducted extensive evaluations using six publicly available benchmark datasets: PAMAP2 [38, 39], DSADS [7, 1], MHealth [5, 6], Realworld [46], UCI-HAR [40, 2], and USCHAD [56]. |
| Dataset Splits | Yes | For fine-tuning and evaluation purposes, we allocate 20% of each dataset for fine-tuning while the remaining data is used for testing. |
| Hardware Specification | Yes | All experiments were conducted on a Quadro RTX 8000 GPU. |
| Software Dependencies | No | The paper mentions software tools like 'Adam W optimizer', 'PyTorch Lightning', and 'Google's text-embedding-004 API', but does not provide specific version numbers for these or other key software dependencies. |
| Experiment Setup | Yes | We set the segment size to 50 samples for Motion Primitive length, used an internal embedding dimension of 256, and masked 25% of motion primitives as prediction targets for the MAE task. In the second stage, the classification head and transformer layers are fine-tuned on each downstream dataset. ... The codebook size is set to 1024 with an embedding dimension of 256. The architecture consists of 5 standard Transformer Encoder layers with 8 attention heads per layer and an MLP ratio of 1. We use GELU as the activation function throughout the network. Notably, we do not employ any dropout in our model. During training, we apply a masking ratio of 0.25 for the masked modeling objective. For optimization, we employ the Adam W optimizer with a learning rate of 1e-4 and a weight decay coefficient of 1e-5. We use a batch size of 512, implemented with gradient accumulation using Py Torch Lightning to optimize memory usage. |